AI Demand Reprices the Entire Chip Stack AI需求重估整个芯片栈

Foundry pricing, HBM packaging, and restricted accelerators all show AI demand pushing costs and capacity decisions through the supply chain. At the same time, Cerebras and Groq face the harder question of turning differentiated silicon into durable economics. 晶圆代工定价、HBM封装和受限加速器都表明，AI需求正把成本与产能决策传导至整条供应链。与此同时，Cerebras和Groq必须回答更难的问题：如何把差异化硅片转化为可持续的商业模式。

AI & Accelerators AI与加速器

Google taps MediaTek for a unified TPUv9 design Google携手MediaTek打造统一型TPUv9

Wccftech · 6h ago Wccftech · 6小时前

Google is reportedly working with MediaTek on TPUv9 “Triggerfish,” combining a CPU and compute die in one package for both training and inference. The design would broaden MediaTek's role after the inference-focused TPUv8i and reduce the split between Google's training and inference accelerator lines, although no shipment date or performance data were disclosed. 据报道，Google正与MediaTek合作开发TPUv9「Triggerfish」，在同一封装内集成CPU与计算裸片，同时面向训练和推理。继侧重推理的TPUv8i之后，这一设计将扩大MediaTek在Google TPU路线图中的角色，并缩小训练与推理产品线的分野，但报道未披露出货时间或性能数据。

Foundry & Manufacturing 晶圆代工与制造

Report: TSMC price increases extend back to 7nm 报道称台积电涨价范围延伸至7nm

Wccftech · 6h ago Wccftech · 6小时前

TSMC is reportedly raising prices across much of its process portfolio, with increases reaching back to the 7nm node and catching some customers off guard. Broad-based hikes would extend AI-era cost pressure beyond leading-edge accelerators to products built on older nodes, but the report does not specify the increases or their effective dates. 据报道，台积电正在上调多个制程的价格，涨价范围向成熟方向延伸至7nm，令部分客户措手不及。若涨价广泛落地，AI时代的成本压力将从先进加速器扩散到采用较旧制程的产品，但报道未说明涨幅和生效日期。

Report: TSMC cuts 28nm wafer starts by 25% 报道称台积电削减25%的28nm晶圆投片量

Crypto Briefing · 5h ago Crypto Briefing · 5小时前

A report says TSMC has reduced 28nm wafer starts by 25%, potentially opening room for UMC to gain foundry share. This is a mature-node utilization decision rather than a measure of total chip output, but it could still alter availability and pricing for customers that remain on 28nm. 一则报道指台积电已将28nm晶圆投片量削减25%，可能为联电争取更多代工份额创造空间。这是成熟制程的产能利用决策，并不等同于芯片总产量下降，但仍可能改变28nm客户面对的供给与价格。

Memory 存储

JEDEC approves SPHBM4 packaging standard JEDEC批准SPHBM4封装标准

DIGITIMES · 2h ago DIGITIMES · 2小时前

JEDEC has approved the SPHBM4 standard, creating a broader packaging path for HBM4 and improving the prospects for glass substrates. Standardization matters because it can widen the set of interoperable packaging choices as HBM4 stacks become more complex, though the item provides no adoption timetable. JEDEC已批准SPHBM4标准，为HBM4提供更广的封装路径，并提升玻璃基板的应用前景。随着HBM4堆叠日益复杂，标准化有望扩大可互操作的封装选择，但报道未给出采用时间表。

Report: SK hynix slows HBM4 expansion as Rubin outlook weakens 报道称Rubin前景转弱，SK hynix放缓HBM4扩产

Moomoo · 5h ago Moomoo · 5小时前

A Korean-media report cited by Moomoo says Nvidia may cut Rubin production and SK hynix is slowing HBM4 capacity expansion, contributing to a selloff in memory shares. If confirmed, the shift would challenge assumptions of uninterrupted HBM expansion; the item includes neither company confirmation nor revised production volumes. Moomoo援引韩国媒体称，Nvidia可能削减Rubin产量，SK hynix则在放缓HBM4扩产，由此引发存储股下跌。若获证实，这一变化将挑战HBM产能持续快速扩张的预期；但该报道既无公司确认，也未提供调整后的产量数据。

Equipment & Materials 设备与材料

ASML joins TNO's Eindhoven photonics factory ASML加入TNO埃因霍温光子工厂

IO+ · 4h ago IO+ · 4小时前

ASML has joined TNO's photonics factory in Eindhoven, extending its participation into a shared development environment for integrated photonics. The move matters because photonic process integration and manufacturing tools are moving closer to industrial qualification, although the item gives no investment or capacity figures. ASML已加入TNO位于埃因霍温的光子工厂，进入集成光子学的共享开发环境。此举表明光子工艺集成与制造工具正进一步接近产业化验证，但报道未披露投资额或产能数据。

Policy & Geopolitics 政策与地缘政治

Restricted Nvidia chips double in price on China's black market 受限Nvidia芯片在中国黑市价格翻倍

Financial Times · 59m ago Financial Times · 59分钟前

Nvidia AI accelerators covered by U.S. export restrictions are reportedly selling for twice their original price on China's black market. The premium shows that controls are constraining lawful supply without eliminating demand, increasing both procurement costs and incentives for diversion. 据报道，受美国出口限制的Nvidia AI加速器在中国黑市的售价已达到原价的两倍。这一溢价表明管制压缩了合规供应，却没有消除需求，同时推高了采购成本与转运套利的激励。

The Netherlands joins the U.S. Pax Silica alliance 荷兰加入美国Pax Silica联盟

아시아경제 · 4h ago 아시아경제 · 4小时前

The Netherlands has joined the U.S.-led AI alliance Pax Silica despite continuing disputes involving ASML. Dutch participation brings the country that anchors advanced lithography into a wider supply-chain coordination framework, giving future equipment-policy discussions greater industrial weight. 尽管围绕ASML的争议仍在持续，荷兰仍加入了美国主导的AI联盟Pax Silica。荷兰掌握先进光刻产业的关键环节，其参与使这一供应链协调框架更具产业分量，也将影响后续设备政策讨论。

Earnings & Markets 财报与市场

Besi raises long-term targets on AI packaging demand AI封装需求推动Besi上调长期目标

The Wall Street Journal · 4h ago The Wall Street Journal · 4小时前

BE Semiconductor Industries raised its long-term revenue and profitability targets, citing an AI-driven demand boost. The reset signals that management expects advanced-packaging demand to persist beyond the current cycle, although the available item does not state the new target values or timetable. BE Semiconductor Industries以AI需求上升为由，上调了长期营收与盈利能力目标。这次调整表明管理层预计先进封装需求将延续至本轮周期之后，但现有报道未列出新目标数值或时间表。

Cerebras revenue rises 92% as its quarterly loss reaches $14 million Cerebras营收增长92%，季度净亏损达1400万美元

GamesBeat · 3h ago GamesBeat · 3小时前

Cerebras reported 92% revenue growth and a $14 million net loss in its first quarterly report after going public. The growth demonstrates demand for its systems, but the continuing loss keeps attention on margins and the cost of scaling a wafer-scale computing business. Cerebras在上市后的首份季度财报中录得92%的营收增长，同时净亏损1400万美元。增长证明其系统存在需求，但持续亏损使市场继续关注利润率，以及扩张晶圆级计算业务所需的成本。

Challengers & New Silicon 新势力与新硅

Cerebras reaches a multi-year agreement with OpenAI Cerebras与OpenAI达成多年协议

Bitget · 3h ago Bitget · 3小时前

An SEC filing indicates that Cerebras has reached a multi-year agreement with OpenAI. A named, production-scale customer is an important validation for Cerebras' wafer-scale platform, but the available report gives no contract value, deployment volume, or service scope. 一份SEC申报文件显示，Cerebras已与OpenAI达成多年协议。具名的大规模客户是对Cerebras晶圆级平台的重要验证，但现有报道未披露合同金额、部署规模或服务范围。

Groq raises $650 million to emphasize inference cloud services Groq融资6.5亿美元，重心转向推理云服务

Wowtale · 6h ago Wowtale · 6小时前

Groq reportedly raised $650 million as it shifts emphasis from selling chips toward operating an AI inference cloud. The funding gives Groq more capacity to deploy its own hardware, but the model also raises capital requirements and makes service utilization—not benchmark speed—the key economic test. 据报道，Groq融资6.5亿美元，并把业务重心从芯片销售转向运营AI推理云。新资金将增强其部署自有硬件的能力，但这一模式也提高了资本需求，使服务利用率而非基准速度成为关键的商业检验。